Privacy-preserving culvert predictive models: A federated learning approach

Abstract

In recent years, transportation agencies have shown a growing interest in the adoption of machine learning (ML) models as a means to improve the efficiency of their infrastructure asset management practices. Their limited inventory can, however, hinder the development of robust ML models. Moreover, transportation agencies may be reluctant to share their raw inventory data with others. To address these challenges, we employed a new ML paradigm, federated learning (FL), and demonstrated its potential in the case study of Utah culvert management. The Utah Department of Transportation (UDOT) only owned a limited culvert inventory, but we were able to obtain additional data from five other state DOTs’ inventories. Rather than employing a conventional centralized ML method, we opted for FL. This approach ensured that while UDOT did not directly access raw data from five other DOTs, it could still benefit from their combined insights. The FL-based model with 80.4 % accuracy showed a promising performance between the two centralized learning models developed based on the consolidated and the Utah-only datasets. The federated approach demonstrates that it is possible to develop high-performance ML models without compromising data confidentiality by centralizing data. This pioneering approach sets a precedent for other transportation agencies, suggesting they can navigate data scarcity and privacy concerns while reaping the benefits of collective knowledge.